Power load forecasting plays a critical role in ensuring the reliable and efficient operation of energy systems, supporting economic optimization and sustainable energy management. However, achieving high accuracy and strong scalability in forecasting models remains challenging. To address these issues, this paper proposes a power load forecasting framework that integrates multiple deep learning models. First, the Boruta algorithm is applied to perform comprehensive feature selection. Then, the original load data is decomposed using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN), and the resulting components are further divided into high- and low-frequency signals based on Sample Entropy, enabling more precise processing. For high-frequency components, Variational Mode Decomposition(VMD) is introduced for secondary decomposition to reduce complexity. Next, an Depth-separable Integrable Spatial Channel Attention Network(DSCAN) encoder enhances spatial feature extraction and optimizes channel utilization, while an Adaptive Wavelet Convolutional Network (AWC) combined with an Attention mechanism strengthens the model’s ability to capture meaningful information. Finally, a Temporal Convolutional Network (TCN) serves as the decoder to achieve accurate load forecasting. Case studies on datasets from two regions yield R2 values of 0.995 and 0.992, demonstrating that the proposed method achieves excellent forecasting performance.
Zhang et al. (Mon,) studied this question.